Retinoblastoma Detection via Image Processing and Interpretable Artificial Intelligence Techniques

Retinoblastoma (RB) is a treatable ocular melanoma that is diagnosed early and subsequently cured in the United States but has a poor prognosis in low- and middle-income countries (LMICs). This study is an approach to diagnosing RB in LMICs. Transfer learning methods were utilized to detect RB from fundus imaging. One hundred and forty RB+ and 140 RB- images were acquired from a previous deep-learning study. Then, five models were used: VGG16, VGG19, Xception, Inception v3, and ResNet50 to train them on the dataset. To evaluate these models, we use two metrics that are considered excellent: Dice Similarity Coefficient (DSC) and Intersection-over-Union (IoU). Explainable AI techniques such as SHAP and LIME were implemented into VGG16 and VGG19 to increase the transparency of their decision-making frameworks, which is critical for the use of AI in medicine. We will show that VGG16 is the best at identifying RB, though the other models achieved great levels of prediction. SHAP values typically ranged from -0.06 to 0.10. Transfer learning methods were effective at identifying RB, and explainable AI made the models more viable for clinical settings.